# Yann LeCun and Brown researchers unveil counterfactual world models for robots > Source: > Published: 2026-11-09 20:59:16+00:00 # Yann LeCun and Brown researchers unveil counterfactual world models for robots Photo: Magda Ehlers / Pexels A new pre-print shows how a simple subtraction step could help AI systems reason about actions they never saw Robots are good at doing what they have seen before. Ask them to handle something new and things tend to go sideways. A team of machine learning researchers thinks it has found a cleaner way to close that gap. Their new technique helps world models, the internal simulators used to train physical AI, predict how objects would behave under actions the model never actually observed. The work comes from Randall Balestriero of Brown University, Kerda Varaku, and Yann LeCun of NYU. It is laid out in a pre-print paper titled **“Counterfactual World Models: Natura Non Facit Saltus”**, published on November 9, 2026. ## What a world model actually does Think of a world model as the AI version of the voice in your head that says “if I push this glass, it falls off the table.” It forecasts the consequences of an action before the action happens. That capability matters most for physical AI systems like robots. A robot that can anticipate outcomes can plan, rather than flailing through trial and error in the real world. The catch is that world models learn from what they observe. If a particular action never shows up in the training data, the model has little to go on when asked what it would do. That is where counterfactuals come in. A counterfactual is a “what if” question: what would have happened if the robot had done something different? ## The trick: subtract the surprise The researchers’ core idea is surprisingly compact. The method looks at the discrepancy between what the world model predicted and what actually happened in the real world, then subtracts it out. That leftover difference is treated as a signal. According to the research, it reveals hidden factors that the model could not see directly. The paper’s approach reportedly uses that kind of gap to enable zero-shot inference of physics. Zero-shot means the system can make the inference without being specifically trained on that scenario first. ### AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. The research describes the technique as aligned with the principles of causal inference, the branch of statistics concerned with cause and effect. Causal counterfactual reasoning is usually a multi-step affair, and this method reportedly compresses it into a single subtraction operation. Fewer moving parts means fewer places for things to break. The research says that simplicity supports better debugging, reasoning, and adaptability during testing. The paper’s title nods to this philosophy. “Natura non facit saltus” is a Latin phrase meaning nature does not make jumps, a fitting motto for a method built on modeling physical continuity. ## The numbers from simulation The researchers tested the method in simulated environments where robots had to adapt to unexpected scenarios. The results were not subtle. On their first retry, robots using the counterfactual approach achieved success rates of 35-69%. Baseline methods managed just 14-23% under the same conditions. When robots were allowed to attempt the same task multiple times, success rates climbed further, landing between 51% and 97%. ## Built on parts already on the shelf One of the more practical features of the work is how little new machinery it requires. The approach builds on JEPA, short for Joint-Embedding Predictive Architecture, an established family of models. According to the research, it can be implemented with minimal modifications. Teams can plug in readily available encoders such as DINOv2 or MAE, paired with a small predictor. An encoder is the component that turns raw input, like camera images, into a compressed representation the model can work with. Using off-the-shelf encoders means researchers do not have to build that layer from scratch. ## Where this fits in the broader research race Counterfactual modeling is not a brand-new field. The research notes that related studies on the topic have appeared across 2025 and 2026. What distinguishes this paper, according to the research, is its focus on practical applications in robotics. Plenty of counterfactual work stays theoretical, and this one puts robots through simulated tasks and reports concrete success rates. LeCun’s involvement adds weight. JEPA-style architectures, which predict in an abstract representation space rather than reconstructing every pixel, are closely associated with his vision for how machines should learn about the world. ## What this means for physical AI There are caveats. This is a pre-print, meaning it has not necessarily gone through formal peer review, and all reported results come from simulation. The broad success ranges also hint that the method shines on some tasks far more than others. For now, the takeaway for the research community is practical. If the results hold up, teams already working with JEPA-style models and encoders like DINOv2 or MAE may be able to add counterfactual reasoning without rebuilding their stack. **Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).